Demographic biases in generative image models: analysis and mitigation through prompt engineering
Year: 2026
Type: dissertation
Abstract: Generative text-to-image models often replicate the societal biases present in their training data. This issue is particularly evident in occupational representations, where models tend to show stereotyped demographic profiles linked to the perceived status of a given profession. This study quantifies demographic biases (race, gender, and age) in four open-weight generative models using statistical metrics like ENS and Cramer’s V, focusing on occupational status. To mitigate these imbalances, we introduce a context-aware, LLM-based prompt-rewriting framework that dynamically injects explicit demographic traits to improve representational parity in AI-generated imagery. For binary demographic attributes, the combination of uniform random sampling and sufficient image volume mathematically drives gender stereotypical bias to near-zero across all evaluated models. For higher-cardinality attributes, the framework achieves substantial racial representational gains, with isolated local overcorrections attributable to sampling variance rather than systematic router failure.
Author: Urdiroz Arroyo, Iker
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Reduced inequalities
Open Access status: green
OpenAlex ID: W7222856107